When AI Finances Its Own Demand
The market can be real and economically overheated at the same time
Artificial intelligence is an expensive business. That has been clear from the beginning. Developing models, purchasing chips, building data centers, securing energy, and providing cooling require capital on a scale that only a handful of companies can support.
Recently, however, something even more interesting than the sheer size of these investments has begun to emerge: the way they are being financed.
The companies building the technology are no longer simply selling it. They are investing in their customers, guaranteeing the financing of their infrastructure, entering into long-term leases, and participating in the creation of the very structures that will ultimately purchase their own products.
Put simply, AI suppliers are beginning to finance demand for AI.
This does not necessarily mean that anything fraudulent is taking place. It does not mean that the sales are fictitious or that the technology has no real value. But it does raise a critical question:
When a supplier finances the customer who will purchase its products, to what extent do those sales remain independent evidence of genuine and economically sustainable demand?
The deals changing the picture
AMD has officially announced that it will invest up to $5 billion in Anthropic. At the same time, Anthropic plans to deploy up to two gigawatts of infrastructure based on AMD Instinct GPUs, with the first gigawatt expected during the first half of 2027.
This is not simply an investment in a promising AI company. AMD is investing in an organization that is simultaneously committing to become one of its largest customers, purchasing equipment potentially worth tens of billions of dollars. The investment and the sale are separate transactions, but economically they belong to the same relationship. AMD, Reuters
The reported discussions between Nvidia and OpenAI are taking place on an even larger scale. According to media reports, Nvidia is considering providing guarantees for approximately $250 billion in financing as part of the construction of a 10-gigawatt data center in Ohio. Including the technology required to equip it, the total cost of the project could exceed $500 billion.
These reports concern ongoing discussions, not a finalized agreement. The direction of travel, however, is significant: the world’s largest supplier of AI chips could use its financial strength to make possible the construction of infrastructure that would require enormous quantities of its own products. Reuters
At the same time, Nvidia has reportedly been identified as the party behind long-term leases for a Hut 8 data center in Texas. The initial value of the contracts has been reported at $19.6 billion over fifteen years and could reach approximately $50 billion if renewal options are exercised. The identity of the tenant has not been officially confirmed by the companies involved. If the reports prove accurate, however, this would represent another case in which a chipmaker is not simply waiting for infrastructure to be built, but is actively helping to make that infrastructure financially possible. Reuters
None of these arrangements proves, on its own, that there is a problem. Taken together, however, they show that the boundaries between supplier, investor, financier, customer, and infrastructure operator are becoming increasingly difficult to distinguish.
The cycle of financed demand
The economic chain works roughly like this:
A chip manufacturer invests in an AI company or guarantees its financing. The AI company commits to purchasing enormous amounts of computing capacity. Data center and cloud infrastructure companies use those commitments to secure loans and investment capital. A substantial proportion of the money raised is then spent on chips and systems supplied by the original manufacturer.
The supplier’s sales increase and are presented as evidence that demand for AI remains explosive. Its stronger financial position and higher valuation then allow it to finance the next customer and the next project.
The transactions are real. The chips are manufactured, the data centers are built, and computational workloads are performed. We are not talking about imaginary sales.
We are dealing with something more complex: different points along the same economic chain are often presented as independent confirmations of demand, even though they finance and reinforce one another.
If we removed strategic investments, guarantees, long-term commitments, borrowing, and promises of future purchases from the equation, how much of today’s demand for AI infrastructure could already stand on its own?
That is the question that neither the scale of investment nor the number of chips sold can answer.
Why this is not automatically a “bubble”
Suppliers financing their own customers is not a new practice. It has been used for decades in aviation, telecommunications, energy, and heavy industry.
A manufacturer may have a stronger balance sheet than its customers and help them overcome the initial cost of an investment. If the resulting activity generates sufficient revenue, the customer repays the financing, the supplier records genuine sales, and the market expands.
The same could happen in AI.
The usage is real. Millions of people use AI models every day. Businesses are integrating AI tools into their operations. APIs are serving a growing number of applications. Training models requires substantial computing resources, while wider adoption is also increasing the demand for inference.
This is not, therefore, a market without products or customers.
The existence of genuine usage, however, does not by itself prove the economic viability of the infrastructure being built to support it.
A product can achieve enormous adoption while still being offered at a price below its true cost. It can generate social, scientific, or operational value without producing enough corresponding revenue to repay hundreds of billions of dollars invested in infrastructure.
Technological utility is one thing. Commercial demand is another. The ability of that demand to finance the entire system supporting it is something else again.
The question is not whether AI will succeed
The public debate is often trapped between two extremes.
On one side are those who regard every new investment as confirmation that AI will transform everything and that the scale of the infrastructure itself proves its inevitable success.
On the other are those who treat every sign of excess as evidence that the entire AI industry is a bubble that will eventually collapse.
Reality may be far less convenient for both sides.
AI may become one of the most important technological developments of our time while, at the same time, more infrastructure is being built than can currently be justified economically.
The internet permanently changed the world, but that did not prevent thousands of companies from collapsing during the dot-com era. Railways transformed economies, but large parts of the networks built around them proved to be poor investments. A technology can win historically while many of those who financed its excessive expansion lose financially.
AI does not have to fail for some of today’s investments to prove excessive.
It is enough for actual revenue from businesses and consumers to grow too slowly to cover the construction of data centers, energy costs, cooling, borrowing, operations, maintenance, and the continuous replacement of technological equipment.
Infrastructure built for decades, technology built for years
There is another contradiction that should not be ignored.
Data centers, energy projects, and transmission networks are financed over many years, sometimes even decades. The chips installed inside them have a much shorter economic life cycle.
A new generation of processors can dramatically alter performance, energy consumption, and the cost of each computational task. New architectures or more efficient models could reduce the amount of computing power required to achieve the same result. Lower inference costs may increase overall usage, but they may also change the economic value of infrastructure designed around today’s prices and technical requirements.
The buildings and energy projects will still exist. The question is whether the equipment and economic assumptions on which their financing was based will remain equally competitive.
The risk is not only insufficient demand. It is also technological obsolescence before the investment has been financially recovered.
Who ultimately carries the risk?
As long as AI companies, chipmakers, and major cloud providers fund these investments with their own capital, the risk remains relatively visible.
As investments are transferred to special financing vehicles, operators, funds, banks, bond markets, and long-term guarantees, however, the risk becomes more widely distributed. Technological expansion remains in the foreground, while the financial obligations are spread across more organizations and over a much longer period.
This may make expansion faster. It may also make the system’s true exposure more difficult to assess.
Who will absorb the cost if an AI company fails to generate the expected revenue? What happens if a long-term lease can no longer be serviced? What is a specialized data center worth if its primary customer reduces its requirements? And how independent is the assessment of risk when the supplier benefiting from the construction of the infrastructure is also guaranteeing its financing?
The answers are not yet known. But the fact that they are unknown is not a reason to ignore the questions.
A market does not have to be fake to become overheated
The point is not to label AI a “bubble.” Such a description would be premature, simplistic, and probably inaccurate as a general characterization of a technology that already has significant real-world applications.
The point is to distinguish between three different things:
The technological value of AI.
The genuine use of its products.
And the economic sustainability of the infrastructure being built around it.
These three elements are connected, but they are not identical.
AI can be useful without every application being profitable. It can achieve impressive growth in adoption without current prices covering its total cost. And it can represent the future of technology without every data center, financing structure, or investment valuation created today proving sustainable.
When suppliers invest in their customers, demand does not stop being real. It does, however, stop being entirely independent of the companies presenting it as evidence of market growth.
That is where greater transparency and more serious evaluation are needed.
Not only how many chips were sold.
Not only how many gigawatts were announced.
Not only how many billions were committed.
But how much genuine economic value is being created at the end of the chain — and whether that value is sufficient for the entire ecosystem to keep operating without requiring a constant flow of new financing from the same participants.
Because the critical question is not whether artificial intelligence creates demand.
It is whether it already creates enough independent value to pay for the enormous structure being built in its name.